slowlp
← loop
Lesson 2026.07.02 · 10 min read

Claude Code's secret China-detection code exposed; GPT-5.6 Pro lineup leaked in genomics paper

Today's AI news — 2026-07-02

LOOP

📰 News

Claude Code secretly flagged Chinese users — and got caught
Since version 2.1.91, Claude Code has been detecting users in China by checking their timezone and proxy URL, then embedding that signal invisibly into system prompts using steganography (swapping a subtly different apostrophe character). The code was obfuscated with XOR encryption and left out of release notes. After a Reddit post exposed it, Anthropic called it “an experiment to prevent account abuse from unauthorized resellers” and announced a rollback. — The Decoder

Trump lifts export restrictions on Anthropic’s Mythos and Fable
The Commerce Department reversed its June 12 export ban on Anthropic’s most powerful models. The pressure point: Asian AI companies began releasing models approaching Mythos-level capabilities, raising concerns about US competitiveness. Anthropic agreed to proactively detect security risks and cooperate with the government. Access restoration began July 1. — TechCrunch

Sonnet 5: same token price, but costs more per task than Opus 4.8
Artificial Analysis found the average task cost with Sonnet 5 at $2.29 — higher than Opus 4.8 at $1.97 — despite identical nominal token rates ($3/M input). The culprit: Sonnet 5 runs up to 3x more agentic loops than its predecessor, and the new tokenizer uses ~30–40% more tokens on the same English text. The per-task cost nearly doubled vs. Sonnet 4.6. — The Decoder

OpenAI’s genomics paper accidentally revealed a three-tier Pro lineup
A benchmark results table listed “GPT-5.6 Luna Pro,” “Terra Pro,” and “Sol Pro” — the first time Pro has ever appeared as more than a single top-tier model. Sol Pro hit a 31.5% pass rate, beating every other model in the 60-model field, and more than doubling Claude Opus 4.8’s score. Whether these variants will actually ship in ChatGPT is still unclear. — The Decoder

Gemini Spark lands on Mac
Google added its agentic assistant to the Mac desktop app. Spark can sort local files, connect to Canva, Instacart, OpenTable, and Zillow Rentals, and track topics in real time (stocks, scores, news). Custom MCP support is coming. Currently in beta for US AI Ultra subscribers only. — TechCrunch

Meta is building a cloud business to sell spare AI compute
Taking a page from SpaceX’s playbook, Meta plans to rent out its surplus GPU capacity to outside customers. Stock jumped 10% on the news. SpaceX already earns $1.25B/month from Anthropic and $920M/month from Google on similar arrangements. Meta’s move raises an implicit question: if there’s GPU capacity to spare, why is internal model development not absorbing it? — The Decoder

  • browser-use/video-use — Drop raw footage in a folder, chat with Claude Code, get a finished video back. Handles cuts, captions, color grading, and animation overlays. Fully open source.
  • usestrix/strix — Autonomous AI pentest agents that run live exploits and generate working proof-of-concepts, not just static analysis reports. CI/CD pipeline integration available.
  • diegosouzapw/OmniRoute — One endpoint to route Claude Code, Cursor, Codex, and Copilot through 231+ AI providers (50+ free), with 15–95% token compression to stretch free-tier limits further.
  • xbtlin/ai-berkshire — Value investing research framework for Claude Code implementing four masters’ methodologies (Buffett, Munger, Duan Yongping, Li Lu) as multi-agent adversarial analysis.

▶️ YouTube

  • AI Just Entered A New Era (Two Minute Papers) — Covers GLM-5.2’s benchmark surge and argues the open-source competitive landscape has fundamentally shifted against closed models.

💬 Community

COMMENTS